CAFES-QM9: A Molecular Dataset with DFT and CCSD Electronic Properties

This dataset contains 10,000 molecular geometries from 500 QM9 molecules, with 20 configurations per molecule. Molecules were selected using size-stratified molecular-fingerprint diversity sampling. Configurations were obtained from B3LYP/def2-SVP ab initio molecular dynamics and selected by potential-energy stratification over the 200–1,000 fs trajectory window.

Each configuration includes atomic numbers and Cartesian coordinates; DFT, Hartree–Fock (HF), and coupled-cluster singles and doubles (CCSD) total energies; DFT and HF atomic-orbital Fock matrices; overlap matrices; the spin-summed, unrelaxed CCSD one-particle density matrix; HF molecular-orbital coefficients, energies, and occupations; CCSD singles amplitudes; and scalar T1 and D1 diagnostics. CCSD calculations used PySCF 2.13.1 with an all-electron restricted HF reference and the def2-SVP basis, without frozen-core approximations, density fitting, or perturbative triples.

The dataset is provided as five independently readable HDF5 files, each containing 100 molecules and 2,000 configurations, with one group per configuration. Original numerical precision and scientific provenance are preserved. No predefined training, validation, or test partitions are provided. The package includes a Python reader, field documentation, a sample index, and integrity checksums. It supports studies of molecular electronic structure and electron correlation across molecular geometries. The README documents a known discrepancy in the recorded MD SCF iteration limits; the numerical data remain unchanged.

Related paper and citation

If you use this dataset, please cite this Zenodo record and the associated CAFES paper:

Tenghui Li and Chong Sun. Automated Many-Body Simulations of Strongly Correlated Systems Using a Correlation-Aware Agentic Framework (2026). arXiv:2610.00943.

@misc{li2026cafes,
  title={Automated Many-Body Simulations of Strongly Correlated Systems Using a Correlation-Aware Agentic Framework},
  author={Tenghui Li and Chong Sun},
  year={2026},
  eprint={2610.00943},
  archivePrefix={arXiv},
  primaryClass={physics.comp-ph},
  url={https://arxiv.org/abs/2610.00943},
}